Improve efficiency of minimum/maximum(::Diagonal)

[jlapeyre]

julia> DM = Diagonal(rand(100));

julia> @btime minimum($DM);    # before
  27.987 μs (0 allocations: 0 bytes)

julia> @btime minimum($DM);    # after
  246.091 ns (0 allocations: 0 bytes)

[mcognetta]

Should maximum be included also?

[jlapeyre]

Should maximum be included also?

Yes, it should. I'll push another commit to this PR.

I recall checking maximum and finding that there was no performance gain. Puzzling, it didn't make sense. I just tried again now, and indeed the same kind of method yields the same efficiency gain.

[mcognetta]

On this note, these methods can also be sped up for the other diagonal matrix types. In my opinion, one nice way to do it would be to write an iterator that chains the diagonals of the matrices. Then you could just find the min/Max over that iterator. The iterator could be used for some other methods like norm.

[andreasnoack]

Maybe make it slightly more general with something like

function Base.mapreduce(f, op::Union{typeof(min),typeof(max)}, D::Diagonal{T}) where T
    rdiag = mapreduce(f, op, diag(D))
    return size(D, 1) > 1 ? op(f(zero(T)), rdiag) : rdiag
end

since minimum and maximum will end up calling this method.

[jlapeyre]

Maybe make it slightly more general with something like

Yes. I'll look into this.

[cossio]

Bump

[tkf]

We need to fix #36287 before merging this to ensure isequal(maximum(d), maximum(collect(d))).

[ViralBShah]

Since #36287 is fixed, should we go ahead and merge this?